Doppler-shift invariant feature extraction for underwater acoustic target classification

Spectrum is one of the commonly used but effective feature for underwater acoustic target classification. However, features extracted based on the spectra are vulnerable to the change of acoustic channels. We propose in this paper a Doppler-shift invariant spectrum feature extraction method which can extract inherent and stable features when target moves in a highly maneuverable manner. The signal is firstly filtered by a set of quasi-orthogonal triangular filters into its frequency domain with a linear property in the log scale. Then the time-variant Doppler shift for each prominent line component tends to be a frequency-invariant offset, which can be easily removed. Such a Doppler-shift invariant feature removes the irrelevant target's motion information and separates out the inherent target spectrum feature. Numerical real data results demonstrate that the proposed feature outperforms the traditional ones in underwater target classification.